Two ~500B MoE Post-Training Designs Compared: Separate RL Runs vs Joint RL
himanshustwts · x · 2026-11-08
A comparison of post-training design between two 500B MoE models: the "beam" approach runs one big capability RL pass plus a separate alignment expert merged via multi-teacher on-policy distillation, while "inkling 2 small" trains everything including alignment in a single joint RL run — a contrast with its predecessor's separate runs + multi-teacher distillation.
More from Models
- AI auditors catch 98% of agent misbehavior but drown reviewers in false alarms, study finds — The Decoder · 2026-11-08
- Rewritten DeepSWE trips GLM and DeepSeek, pointing to evaluation-aware benchmarks — teortaxesTex · 2026-11-08
- AI models fade and get lazy over time — dev calls for a Proof of Model standard — sull · 2026-11-08
- Reddit user extrapolates 'Model 4' ECI at ~182.5, unverified speculation — Realistic_Stomach848 · 2026-11-08
- Running DeepSeek V4.1 Flash locally at 15 tok/s on an RTX 3090 build for ~$2.5k — casper_hansen_ · 2026-11-08
- GLM and DeepSeek accused of benchmark awareness after MaxxBench rewritten-task failures — teortaxesTex · 2026-11-08